Presented by

Arca logo

What's Actually Happening

A week after Anthropic took the lead with Claude Fable 5.2, OpenAI answered. GPT-6.2 Astra and GPT-6.2 Sol landed today in ChatGPT and the API, and the headline isn't raw smarts. It's price: OpenAI says Astra sets new highs at less than a third of the cost per task of the model it replaces.

It also comes with a first. OpenAI says Astra is the first model it has released that reaches its "High" risk level for AI self-improvement.

Also tonight: Microsoft launches seven in-house MAI models aimed squarely at OpenAI and Anthropic pricing, California sues OpenAI over rogue agent incidents, and four tools worth trying this week.

OpenAI released GPT-6.2 Astra, its new frontier model for long, complex projects, and GPT-6.2 Sol, a cheaper workhorse, on November 17. Both are rolling out now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol. They aren't in regular ChatGPT chat yet.

In the API, Astra costs $8 per million input tokens and $40 per million output, which The Decoder reports is 20% cheaper than GPT-6 Astra; Sol stays at $2 and $10. On Vals AI's cost-per-task measure, Astra comes in at $10.68 against $21.28 for GPT-6 Astra, and Sol at $3.18. OpenAI says Astra beats Claude Fable 5.2 by 5.5 points on Agents' Last Exam at a fraction of the cost, and Sol nearly matches Astra on real-codebase changes at about a fifth of the price. Pro 500 subscribers also get Ultrafast mode on both models.

Our take: this is a price war move as much as a capability one. Independent results are mixed: Artificial Analysis puts Astra second on its Intelligence Index (62) behind Fable 5.2 (65), Vals ranks Fable first and Astra seventh, and Artificial Analysis says coding has stalled, with Sol a few points below GPT-6.1 Sol. OpenAI also rated both models "Critical" for cybersecurity and Astra "High" for self-improvement, and says tighter cyber safeguards will block some legitimate security work. If you're paying Astra prices for agent work, test Sol first.

Will you try GPT-6.2 Astra or Sol this week?

Login or Subscribe to participate

⚠️ Why It Matters: The Self-Improvement Line Got Crossed

OpenAI's Preparedness Framework rates models on how much they could speed up AI research itself. Astra is the first model OpenAI has shipped that reaches the "High" level there, and OpenAI lists it as a risk. Its answer is tighter controls inside its own research environments and a monitoring system it calls Control Room, rather than holding the model back.

The timing is awkward. A day earlier, California's attorney general sued OpenAI, alleging that a pre-release model broke out of its test environment in July and compromised Hugging Face's systems to finish a benchmark, and citing dozens more incidents. OpenAI says the suit unfairly singles it out. On the same day, it announced that six Apollo Research evaluators will work inside the company for four months to look for models undermining safeguards. Expect every frontier launch from here to come with a safety argument attached.

A Word From Today's Sponsor

Arca logo
Arca by Brevis: get paid for the AI sessions you'd otherwise throw away

If you use Claude Code, Codex or Cursor every day, your computer is full of old sessions that just sit there.

Arca is a free desktop app for Mac and Windows that finds them and pays you for the ones you choose to share.

  • You pick what to share. Nothing uploads in the background.

  • Names, file paths, account details and keys are removed on your computer first.

  • Each session gets a quote range. Submit it for review, or withdraw it and the upload is deleted.

  • Approved sessions are paid weekly.

It's open to everyone now, with no invite or waitlist.

Let’s Talk MODELS! 🤖

Microsoft used Ignite to launch seven of its own MAI models, each aimed at a specific job and priced well under the OpenAI and Anthropic models it usually resells. The headliner is MAI Code 2, now in GitHub Copilot, which Microsoft says reaches 98% of GPT-6.1 Sol's coding performance and 96% of Claude Opus 5.5's at about 40% of Sol's price and about 5% of Opus 5.5's. MAI Code 2 Flash handles quick edits and fixes, and is also in a limited Foundry preview.

The rest of the lineup: MAI Cyber 2 Flash for always-on security agents, MAI Image 2.7 (which Microsoft says ranks first on Artificial Analysis image editing), MAI Realtime 1 for live voice from $0.01 per audio minute, Dragon Copilot Transcribe for medical notes, and Decision 1.1, a router that sends requests to bigger models only when needed.

Our take: Microsoft is building an exit from its dependence on OpenAI, one job at a time. The benchmark and price comparisons are Microsoft's own, and it says MAI Code 2 isn't tuned yet for very long agent runs or hard research problems, but for everyday Copilot work a model this close to Sol at a fraction of the price is hard to ignore.

The Cheapest Model In The Room

Tencent released Hy4 flash on November 16, a multimodal model with 224 billion total parameters, only 12 billion active per token, and up to 1 million tokens of context. Tencent lists it at about $0.07 per million input tokens and $0.28 per million output, its most cost-efficient Hy model yet.

It's live in Tencent Cloud TokenHub (at 50% off for now), the WorkBuddy app (free through November 30), and Tencent's Yuanbao, WeCom and QQ apps, with OpenRouter support coming soon. Tencent says it plans to open-source the weights but hasn't said when.

Our take: next to Astra's $40 output price, this is a different universe. It won't match a frontier model on hard work, but for high-volume summarizing, tagging and routing jobs, prices this low change what's worth automating.

Top 5 In AI Research 🔬

The other stories worth your attention:

California sued OpenAI over rogue agent incidents, centered on a pre-release model that allegedly breached Hugging Face's systems to finish a benchmark in July. See the other sites the complaint names. (TechCrunch)

OpenAI is letting outside evaluators work inside the company, starting with six Apollo Research staff hunting for models that undermine safeguards. See which other safety groups signed on. (OpenAI)

The US tightened AI chip export rules to stop Chinese firms renting restricted chips remotely through overseas data centers. See which region the rule targets. (Reuters)

Google caught an AI hacking framework turning security patches into working exploits with no human help. See who Google links it to. (The Register)

Perplexity launched a frontier research lab to build AI that forms hypotheses and runs its own experiments, and plans to publish its code and models openly. See how many GPUs are behind it. (Perplexity)

🛠️ Tools That Are Hot Right Now!

  • 🤗 Hugging Face x OpenRouter - call almost any open model on the Hub with one Hugging Face token, routed to the cheapest or fastest provider automatically, with setup guides for Claude Code and Codex. No markup on provider prices; see what open models cost on average versus closed ones.

  • 🗺️ Ask Maps itineraries - have Google Maps plan a multi-day trip with live details for every stop, then reshuffle it mid-trip. Free in Google Maps; see what else it can book for you.

  • 🧩 CopilotKit - build the user-facing interface for your AI agent, and host it yourself. See how it plugs into the agent framework you already use.

  • 🔐 OpenAI's child-safety image classifier - detect potential abuse imagery, including AI-generated images, on your platform. Free via API for vetted organizations; see who qualifies.

What's The Recap?

OpenAI answered Claude Fable 5.2 with GPT-6.2 Astra and Sol, leading on price per task more than on independent leaderboards, and flagged Astra as its first model to reach "High" for self-improvement risk. Microsoft launched seven MAI models priced to undercut the labs it partners with, Tencent's Hy4 flash went to 7 cents per million input tokens, and California took OpenAI to court over rogue agents. The thread: the frontier is now competing on cost per job, and the safety questions are getting harder to set aside.

Login or Subscribe to participate

Stay building. 🤖

Our newest video is live. Watch it below.

📬 Read by 14K AI readers. Know someone who'd like it? Send them here to subscribe.

Recommended for you

View all
caret-right